I tried to train ENet from scratch on my own data and after several epochs of training ENet started to return a tensor with Nan values. Can any part of the model cause this problem? If it is true, how we can handle it?
Something like that is more likely to be connected to faulty input images (large values, NaN, Inf) or training parameters that cause instability (high learning rate).
I tried to train ENet from scratch on my own data and after several epochs of training ENet started to return a tensor with Nan values. Can any part of the model cause this problem? If it is true, how we can handle it?